1. Parallel & Cluster Computing 2005 Distributed Multiprocessing and MPI
National Computational Science Institute
July 31 – August
Paul Gray, University of Northern Iowa
David Joiner, Kean University
Tom Murphy, Contra Costa College
Henry Neeman, University of Oklahoma
Charlie Peck, Earlham College

2. Outline What is a Cluster? The Desert Islands Analogy
Distributed Parallelism
MPI
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

3. What is a Cluster?

4. What is a Cluster?
“… [W]hat a ship is … It's not just a keel and hull and a deck and sails. That's what a ship needs. But what a ship is... is freedom.”
– Captain Jack Sparrow
“Pirates of the Caribbean”
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

5. What a Cluster is ….
A cluster needs of a collection of small computers, called nodes, hooked together by an interconnection network (or interconnect for short).
It also needs software that allows the nodes to communicate over the interconnect.
But what a cluster is … is the relationships among these components.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

6. What’s in a Cluster?
These days, many clusters are made of nodes that are basically Linux PCs, made out of the same hardware components as you’d find in your own PC.
As for interconnects, it’s quite common for clusters these days to use regular old Ethernet (typically Gigabit), but there are also several high performance interconnects on the market:
Dolphin Wulfkit (SCI)
Infiniband
Myrinet
Quadrics
10 Gigabit Ethernet
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

7. An Actual Cluster boomer.oscer.ou.edu Interconnect Nodes
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

8. Details about boomer 270 Pentium4 Xeon 2 GHz CPUs 270 GB RAM
8700 GB disk
OS: Red Hat Linux Enterprise 3.0
Peak speed: 1.08 TFLOP/s*
Programming model:
distributed multiprocessing
*TFLOP/s: trillion floating point calculations per second
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

9. More boomer Details Each node has Breakdown of nodes
2 Pentium4 XeonDP CPUs (2.0 GHz)
2 GB RAM
1 or more hard drives (mostly EIDE, a few SCSI)
100 Mbps Ethernet (cheap backup interconnect)
Myrinet2000 (high performance interconnect)
Breakdown of nodes
135 compute nodes
6 storage nodes
2 “head” nodes (where you log in, compile, etc)
1 management node (batch system, LDAP, etc)
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

10. The Desert Islands Analogy

11. An Island Hut Imagine you’re on a desert island in a little hut.
Inside the hut is a desk and a chair.
On the desk is:
a phone;
a pencil;
a calculator;
a piece of paper with instructions;
a piece of paper with numbers.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

12. Instructions The instructions are split into two kinds:
Arithmetic/Logical: e.g.,
Add the 27th number to the 239th number
Compare the 96th number to the 118th number to see whether they are equal
Communication: e.g.,
dial and leave a voic containing the 962nd number
call your voic box and collect a voic from and put that number in the 715th slot
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

13. Is There Anybody Out There?
If you’re in a hut on an island, you aren’t specifically aware of anyone else.
Especially, you don’t know whether anyone else is working on the same problem as you are, and you don’t know who’s at the other end of the phone line.
All you know is what to do with the voic s you get, and what phone numbers to send voic s to.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

14. Someone Might Be Out There
Now suppose that Paul is on another island somewhere, in the same kind of hut, with the same kind of equipment.
Suppose that he has the same list of instructions as you, but a different set of numbers (both data and phone numbers).
Like you, he doesn’t know whether there’s anyone else working on his problem.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

15. Even More People Out There
Now suppose that Dave and Tom are also in huts on islands.
Suppose that each of the four has the exact same list of instructions, but different lists of numbers.
And suppose that the phone numbers that people call are each others’. That is, your instructions have you call Paul, Dave and Tom, Paul’s has him call Dave, Tom and you, and so on.
Then you might all be working together on the same problem, even though you’re not aware of it.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

16. All Data Are Private
Notice that you can’t see Paul’s or Dave’s or Tom’s numbers, nor can they see yours or each other’s.
Thus, everyone’s numbers are private: there’s no way for anyone to share numbers, except by leaving them in voic s.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

17. Long Distance Calls: 2 Costs
When you make a long distance phone call, you typically have to pay two costs:
Connection charge: the fixed cost of connecting your phone to someone else’s, even if you’re only connected for a second
Per-minute charge: the cost per minute of talking, once you’re connected
If the connection charge is large, then you want to make as few calls as possible.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

18. Distributed Parallelism

19. Like Desert Islands
Distributed parallelism is very much like the Desert Islands analogy:
processes are independent of each other.
All data are private.
Processes communicate by passing messages (like voic s).
The cost of passing a message is split into:
latency (connection time)
bandwidth (time per byte)
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

20. Parallelism
Parallelism means doing multiple things at the same time: you can get more work done in the same amount of time.
Less fish …
More fish!
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

21. What Is Parallelism?
Parallelism is the use of multiple processing units – either processors or parts of an individual processor – to solve a problem, and in particular the use of multiple processing units operating concurrently on different parts of a problem.
The different parts could be different tasks, or the same task on different pieces of the problem’s data.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

22. Kinds of Parallelism
Shared Memory Multithreading (our topic last time)
Distributed Memory Multiprocessing (today)
Hybrid Shared/Distributed
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

23. Why Parallelism Is Good
The Trees: We like parallelism because, as the number of processing units working on a problem grows, we can solve the same problem in less time.
The Forest: We like parallelism because, as the number of processing units working on a problem grows, we can solve bigger problems.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

24. Parallelism Jargon
Threads: execution sequences that share a single memory area (“address space”)
Processes: execution sequences with their own independent, private memory areas
… and thus:
Multithreading: parallelism via multiple threads
Multiprocessing: parallelism via multiple processes
As a general rule, Shared Memory Parallelism is concerned with threads, and Distributed Parallelism is concerned with processes.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

25. Jargon Alert In principle:
“shared memory parallelism”  “multithreading”
“distributed parallelism”  “multiprocessing”
In practice, these terms are often used interchangeably:
Parallelism
Concurrency (not as popular these days)
Multithreading
Multiprocessing
Typically, you have to figure out what is meant based on the context.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

26. Load Balancing
Suppose you have a distributed parallel code, but one process does 90% of the work, and all the other processes share 10% of the work.
Is it a big win to run on 1000 processes?
Now, suppose that each process gets exactly 1/Np of the work, where Np is the number of processes.
Now is it a big win to run on 1000 processes?
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

27. Load Balancing
Load balancing means giving everyone roughly the same amount of work to do.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

28. Load Balancing
Load balancing can be easy, if the problem splits up into chunks of roughly equal size, with one chunk per process. Or load balancing can be very hard.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

29. Load Balancing Is Good
When every process gets the same amount of work, the job is load balanced.
We like load balancing, because it means that our speedup can potentially be linear: if we run on Np processes, it takes 1/Np as much time as on one.
For some codes, figuring out how to balance the load is trivial (e.g., breaking a big unchanging array into sub-arrays).
For others, load balancing is very tricky (e.g., a dynamically evolving collection of arbitrarily many blocks of arbitrary size).
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

30. Parallel Strategies
Client-Server: One worker (the server) decides what tasks the other workers (clients) will do; e.g., Hello World, Monte Carlo.
Data Parallelism: Each worker does exactly the same tasks on its unique subset of the data; e.g., distributed meshes (weather etc).
Task Parallelism: Each worker does different tasks on exactly the same set of data (each process holds exactly the same data as the others); e.g., N-body.
Pipeline: Each worker does its tasks, then passes its set of data along to the next worker and receives the next set of data from the previous worker.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

31. MPI: The Message-Passing Interface
Most of this discussion is from [1] and [2].

32. What Is MPI?
The Message-Passing Interface (MPI) is a standard for expressing distributed parallelism via message passing.
MPI consists of a header file, a library of routines and a runtime environment.
When you compile a program that has MPI calls in it, your compiler links to a local implementation of MPI, and then you get parallelism; if the MPI library isn’t available, then the compile will fail.
MPI can be used in Fortran, C and C++.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

33. MPI Calls MPI calls in Fortran look like this:
CALL MPI_Funcname(…, errcode)
In C, MPI calls look like:
errcode = MPI_Funcname(…)
In C++, MPI calls look like:
errcode = MPI::Funcname(…)
Notice that errcode is returned by the MPI routine MPI_Funcname, with a value of MPI_SUCCESS indicating that MPI_Funcname has worked correctly.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

34. MPI is an API
MPI is actually just an Application Programming Interface (API).
An API specifies what a call to each routine should look like, and how each routine should behave.
An API does not specify how each routine should be implemented, and sometimes is intentionally vague about certain aspects of a routine’s behavior.
Each platform has its own MPI implementation.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

35. Example MPI Routines
MPI_Init starts up the MPI runtime environment at the beginning of a run.
MPI_Finalize shuts down the MPI runtime environment at the end of a run.
MPI_Comm_size gets the number of processes in a run, Np (typically called just after MPI_Init).
MPI_Comm_rank gets the process ID that the current process uses, which is between 0 and Np-1 inclusive (typically called just after MPI_Init).
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

36. More Example MPI Routines
MPI_Send sends a message from the current process to some other process (the destination).
MPI_Recv receives a message on the current process from some other process (the source).
MPI_Bcast broadcasts a message from one process to all of the others.
MPI_Reduce performs a reduction (e.g., sum, maximum) of a variable on all processes, sending the result to a single process.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

37. MPI Program Structure (F90)
PROGRAM my_mpi_program
IMPLICIT NONE
INCLUDE "mpif.h"
[other includes]
INTEGER :: my_rank, num_procs, mpi_error_code
[other declarations]
CALL MPI_Init(mpi_error_code) !! Start up MPI
CALL MPI_Comm_Rank(my_rank, mpi_error_code)
CALL MPI_Comm_size(num_procs, mpi_error_code)
[actual work goes here]
CALL MPI_Finalize(mpi_error_code) !! Shut down MPI
END PROGRAM my_mpi_program
Note that MPI uses the term “rank” to indicate process identifier.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

38. MPI Program Structure (in C)
#include <stdio.h>
#include "mpi.h"
[other includes]
int main (int argc, char* argv[])
{ /* main */
int my_rank, num_procs, mpi_error;
[other declarations]
mpi_error = MPI_Init(&argc, &argv); /* Start up MPI */
mpi_error = MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);
mpi_error = MPI_Comm_size(MPI_COMM_WORLD, &num_procs);
[actual work goes here]
mpi_error = MPI_Finalize(); /* Shut down MPI */
} /* main */
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

39. Example: Hello World Start the MPI system.
Get the rank and number of processes.
If you’re not the server process:
Create a “hello world” string.
Send it to the server process.
If you are the server process:
For each of the client processes:
Receive its “hello world” string.
Print its “hello world” string.
Shut down the MPI system.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

40. hello_world_mpi.c #include <stdio.h> #include <string.h>
#include "mpi.h"
int main (int argc, char* argv[])
{ /* main */
const int maximum_message_length = 100;
const int server_rank = 0;
char message[maximum_message_length+1];
MPI_Status status; /* Info about receive status */
int my_rank; /* This process ID */
int num_procs; /* Number of processes in run */
int source; /* Process ID to receive from */
int destination; /* Process ID to send to */
int tag = 0; /* Message ID */
int mpi_error; /* Error code for MPI calls */
[work goes here]
} /* main */
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

41. Hello World Startup/Shut Down
[header file includes]
int main (int argc, char* argv[])
{ /* main */
[declarations]
mpi_error = MPI_Init(&argc, &argv);
mpi_error = MPI_Comm_rank(MPI_COMM_WORLD, &my_rank);
mpi_error = MPI_Comm_size(MPI_COMM_WORLD, &num_procs);
if (my_rank != server_rank) {
[work of each non-server (worker) process]
} /* if (my_rank != server_rank) */
else {
[work of server process]
} /* if (my_rank != server_rank)…else */
mpi_error = MPI_Finalize();
} /* main */
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

42. Hello World Client’s Work
[header file includes]
int main (int argc, char* argv[])
{ /* main */
[declarations]
[MPI startup (MPI_Init etc)]
if (my_rank != server_rank) {
sprintf(message, "Greetings from process #%d!“,
my_rank);
destination = server_rank;
mpi_error =
MPI_Send(message, strlen(message) + 1, MPI_CHAR,
destination, tag, MPI_COMM_WORLD);
} /* if (my_rank != server_rank) */
else {
[work of server process]
} /* if (my_rank != server_rank)…else */
mpi_error = MPI_Finalize();
} /* main */
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

43. Hello World Server’s Work
[header file includes]
int main (int argc, char* argv[])
{ /* main */
[declarations, MPI startup]
if (my_rank != server_rank) {
[work of each client process]
} /* if (my_rank != server_rank) */
else {
for (source = 0; source < num_procs; source++) {
if (source != server_rank) {
mpi_error =
MPI_Recv(message, maximum_message_length + 1,
MPI_CHAR, source, tag, MPI_COMM_WORLD,
&status);
fprintf(stderr, "%s\n", message);
} /* if (source != server_rank) */
} /* for source */
} /* if (my_rank != server_rank)…else */
mpi_error = MPI_Finalize();
} /* main */
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

44. How an MPI Run Works
Every process gets a copy of the executable: Single Program, Multiple Data (SPMD).
They all start executing it.
Each looks at its own rank to determine which part of the problem to work on.
Each process works completely independently of the other processes, except when communicating.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

45. Compiling and Running
% mpicc -o hello_world_mpi hello_world_mpi.c
% mpirun -np 1 hello_world_mpi
% mpirun -np 2 hello_world_mpi
Greetings from process #1!
% mpirun -np 3 hello_world_mpi
Greetings from process #2!
% mpirun -np 4 hello_world_mpi
Greetings from process #3!
Note: the compile command and the run command vary from platform to platform.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

46. Why is Rank #0 the server? const int server_rank = 0;
By convention, the server process has rank (process ID) #0. Why?
A run must use at least one process but can use multiple processes.
Process ranks are 0 through Np-1, Np >1 .
Therefore, every MPI run has a process with rank #0.
Note: every MPI run also has a process with rank Np-1, so you could use Np-1 as the server instead of 0 … but no one does.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

47. Why “Rank?” Why does MPI use the term rank to refer to process ID?
In general, a process has an identifier that is assigned by the operating system (e.g., Unix), and that is unrelated to MPI:
% ps
PID TTY TIME CMD
ttyq57 0:01 tcsh
Also, each processor has an identifier, but an MPI run that uses fewer than all processors will use an arbitrary subset.
The rank of an MPI process is neither of these.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

48. Compiling and Running Recall:
% mpicc -o hello_world_mpi hello_world_mpi.c
% mpirun -np 1 hello_world_mpi
% mpirun -np 2 hello_world_mpi
Greetings from process #1!
% mpirun -np 3 hello_world_mpi
Greetings from process #2!
% mpirun -np 4 hello_world_mpi
Greetings from process #3!
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

49. Deterministic Operation?
% mpirun -np 4 hello_world_mpi
Greetings from process #1!
Greetings from process #2!
Greetings from process #3!
The order in which the greetings are printed is deterministic. Why?
for (source = 0; source < num_procs; source++) {
if (source != server_rank) {
mpi_error =
MPI_Recv(message, maximum_message_length + 1,
MPI_CHAR, source, tag, MPI_COMM_WORLD,
&status);
fprintf(stderr, "%s\n", message);
} /* if (source != server_rank) */
} /* for source */
This loop ignores the receive order.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

50. Message = Envelope+Contents
MPI_Send(message, strlen(message) + 1,
MPI_CHAR, destination, tag,
MPI_COMM_WORLD);
When MPI sends a message, it doesn’t just send the contents; it also sends an “envelope” describing the contents:
Size (number of elements of data type)
Data type
Source: rank of sending process
Destination: rank of process to receive
Tag (message ID)
Communicator (e.g., MPI_COMM_WORLD)
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

51. MPI Data Types C
Fortran 90
char
MPI_CHAR
CHARACTER
MPI_CHARACTER
int
MPI_INT
INTEGER
MPI_INTEGER
float
MPI_FLOAT
REAL
MPI_REAL
double
MPI_DOUBLE
DOUBLE PRECISION
MPI_DOUBLE_PRECISION
MPI supports several other data types, but most are variations of these, and probably these are all you’ll use.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

52. Message Tags
for (source = 0; source < num_procs; source++) {
if (source != server_rank) {
mpi_error =
MPI_Recv(message, maximum_message_length + 1,
MPI_CHAR, source, tag, MPI_COMM_WORLD,
&status);
fprintf(stderr, "%s\n", message);
} /* if (source != server_rank) */
} /* for source */
The greetings are printed in deterministic order not because messages are sent and received in order, but because each has a tag (message identifier), and MPI_Recv asks for a specific message (by tag) from a specific source (by rank).
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

53. Communicators
An MPI communicator is a collection of processes that can send messages to each other.
MPI_COMM_WORLD is the default communicator; it contains all of the processes. It’s probably the only one you’ll need.
Some libraries (e.g., PETSc) create special library-only communicators, which can simplify keeping track of message tags.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

54. Broadcasting
What happens if one process has data that everyone else needs to know?
For example, what if the server process needs to send an input value to the others?
CALL MPI_Bcast(length, 1, MPI_INTEGER, &
& source, MPI_COMM_WORLD,
error_code)
Note that MPI_Bcast doesn’t use a tag, and that the call is the same for both the sender and all of the receivers.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

55. Broadcast Example: Setup
PROGRAM broadcast
USE mpi
IMPLICIT NONE
INTEGER,PARAMETER :: server = 0
INTEGER,PARAMETER :: source = server
INTEGER,DIMENSION(:),ALLOCATABLE :: array
INTEGER :: length, memory_status
INTEGER :: num_procs, my_rank, mpi_error_code
CALL MPI_Init(mpi_error_code)
CALL MPI_Comm_rank(MPI_COMM_WORLD, my_rank, &
& mpi_error_code)
CALL MPI_Comm_size(MPI_COMM_WORLD, num_procs, &
[input]
[broadcast]
CALL MPI_Finalize(mpi_error_code)
END PROGRAM broadcast
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

56. Broadcast Example: Input
PROGRAM broadcast
USE mpi
IMPLICIT NONE
INTEGER,PARAMETER :: server = 0
INTEGER,PARAMETER :: source = server
INTEGER,DIMENSION(:),ALLOCATABLE :: array
INTEGER :: length, memory_status
INTEGER :: num_procs, my_rank, mpi_error_code
[MPI startup]
IF (my_rank == server) THEN
OPEN (UNIT=99,FILE="broadcast_in.txt")
READ (99,*) length
CLOSE (UNIT=99)
ALLOCATE(array(length), STAT=memory_status)
array(1:length) = 0
END IF !! (my_rank == server)...ELSE
[broadcast]
CALL MPI_Finalize(mpi_error_code)
END PROGRAM broadcast
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

57. Broadcast Example: Broadcast
PROGRAM broadcast
USE mpi
IMPLICIT NONE
INTEGER,PARAMETER :: server = 0
INTEGER,PARAMETER :: source = server
[other declarations]
[MPI startup and input]
IF (num_procs > 1) THEN
CALL MPI_Bcast(length, 1, MPI_INTEGER, source, &
& MPI_COMM_WORLD, mpi_error_code)
IF (my_rank /= server) THEN
ALLOCATE(array(length), STAT=memory_status)
END IF !! (my_rank /= server)
CALL MPI_Bcast(array, length, MPI_INTEGER, source, &
MPI_COMM_WORLD, mpi_error_code)
WRITE (0,*) my_rank, ": broadcast length = ", length
END IF !! (num_procs > 1)
CALL MPI_Finalize(mpi_error_code)
END PROGRAM broadcast
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

58. Broadcast Compile & Run
% mpif90 -o broadcast broadcast.f90
% mpirun -np 4 broadcast
0 : broadcast length =
1 : broadcast length =
2 : broadcast length =
3 : broadcast length =
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

59. Reductions
A reduction converts an array to a scalar: e.g., sum, product, minimum value, maximum value, Boolean AND, Boolean OR, etc.
Reductions are so common, and so important, that MPI has two routines to handle them:
MPI_Reduce: sends result to a single specified process
MPI_Allreduce: sends result to all processes (and therefore takes longer)
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

60. Reduction Example PROGRAM reduce USE mpi IMPLICIT NONE
INTEGER,PARAMETER :: server = 0
INTEGER :: value, value_sum
INTEGER :: num_procs, my_rank, mpi_error_code
CALL MPI_Init(mpi_error_code)
CALL MPI_Comm_rank(MPI_COMM_WORLD, my_rank, mpi_error_code)
CALL MPI_Comm_size(MPI_COMM_WORLD, num_procs, mpi_error_code)
value_sum = 0
value = my_rank * num_procs
CALL MPI_Reduce(value, value_sum, 1, MPI_INT, MPI_SUM, &
& server, MPI_COMM_WORLD, mpi_error_code)
WRITE (0,*) my_rank, ": reduce value_sum = ", value_sum
CALL MPI_Allreduce(value, value_sum, 1, MPI_INT, MPI_SUM, &
& MPI_COMM_WORLD, mpi_error_code)
WRITE (0,*) my_rank, ": allreduce value_sum = ", value_sum
CALL MPI_Finalize(mpi_error_code)
END PROGRAM reduce
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

61. Compiling and Running % mpif90 -o reduce reduce.f90
% mpirun -np 4 reduce
3 : reduce value_sum = 0
1 : reduce value_sum = 0
2 : reduce value_sum = 0
0 : reduce value_sum = 24
0 : allreduce value_sum = 24
1 : allreduce value_sum = 24
2 : allreduce value_sum = 24
3 : allreduce value_sum = 24
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

62. Why Two Reduction Routines?
MPI has two reduction routines because of the high cost of each communication.
If only one process needs the result, then it doesn’t make sense to pay the cost of sending the result to all processes.
But if all processes need the result, then it may be cheaper to reduce to all processes than to reduce to a single process and then broadcast to all.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

63. Example: Monte Carlo
Monte Carlo methods are approximation methods that randomly generate a large number of examples (realizations) of a phenomenon, and then take the average of the examples’ properties.
When the realizations’ average converges (i.e., doesn’t change substantially if new realizations are generated), then the Monte Carlo simulation stops.
Monte Carlo simulations are sometimes known as embarrassingly parallel.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

64. Serial Monte Carlo
Suppose you have an existing serial Monte Carlo simulation:
PROGRAM monte_carlo
CALL read_input(…)
DO WHILE (average_properties_havent_converged(…))
CALL generate_random_realization(…)
CALL calculate_properties(…)
CALL calculate_average(…)
END DO
END PROGRAM monte_carlo
How would you parallelize this?
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

65. Parallel Monte Carlo PROGRAM monte_carlo [MPI startup]
IF (my_rank == server_rank) THEN
CALL read_input(…)
END IF !! (my_rank == server_rank)
CALL MPI_Bcast(…)
DO WHILE (average_properties_havent_converged(…))
CALL generate_random_realization(…)
CALL calculate_properties(…)
[collect properties]
ELSE !! (my_rank == server_rank)
[send properties]
END IF !! (my_rank == server_rank)…ELSE
CALL calculate_average(…)
END DO !! WHILE (average_properties_havent_converged(…))
[MPI shutdown]
END PROGRAM monte_carlo
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

66. Asynchronous Communication
MPI allows a process to start a send, then go on and do work while the message is in transit.
This is called asynchronous or non-blocking or immediate communication. (Here, “immediate” refers to the fact that the call to the MPI routine returns immediately rather than waiting for the send to complete.)
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

67. Immediate Send Likewise:
CALL MPI_Isend(array, size, MPI_FLOAT, &
& destination, tag, communicator, request, &
& mpi_error_code)
Likewise:
CALL MPI_Irecv(array, size, MPI_FLOAT, &
& source, tag, communicator, request, &
This call starts the send/receive, but the send/receive won’t be complete until:
CALL MPI_Wait(request, status)
What’s the advantage of this?
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

68. Communication Hiding
In between the call to MPI_Isend/Irecv and the call to MPI_Wait, both processes can do work!
If that work takes at least as much time as the communication, then the cost of the communication is effectively zero, since the communication won’t affect how much work gets done.
This is called communication hiding.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

69. Communication Hiding in MC
In our Monte Carlo example, we could use communication hiding by, for instance, sending the properties of each realization asynchronously.
That way, the sending process can start generating a new realization while the old realization’s properties are in transit.
The server process can collect the other processes’ data when it’s done with its realization.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

70. Rule of Thumb for Hiding
When you want to hide communication:
as soon as you calculate the data, send it;
don’t receive it until you need it.
That way, the communication has the maximal amount of time to happen in background (behind the scenes).
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August

71. References
[1] P.S. Pacheco, Parallel Programming with MPI, Morgan Kaufmann
Publishers, 1997.
[2] W. Gropp, E. Lusk and A. Skjellum, Using MPI: Portable Parallel
Programming with the Message-Passing Interface, 2nd ed. MIT
Press, 1999.
NCSI Parallel & Cluster Computing OU
Distributed Multiprocessing and MPI
July 31 – August